Huawei Unveils Atlas 950 SuperPoD: 1,024 Ascend Chips, Claims 6.7x Nvidia's Compute

Huawei publicly showed off real Atlas 950 SuperPoD hardware at the World Artificial Intelligence Conference (WAIC 2026) in Shanghai — 1,024 Ascend NPUs packed into 16 cabinets. The message is unmistakable: under US chip export controls, China is building its own AI compute highway.

Three Key Takeaways

Impressive on paper. The demonstrated 1,024-NPU configuration delivers 1 EFLOPS (FP8) / 2 EFLOPS (FP4), 256 TB of globally addressable unified memory, TB-scale NPU interconnect bandwidth, and 3-microsecond round-trip latency. At full scale — 8,192 NPUs, 160 cabinets, roughly 1,000 square meters — the system promises 8 EFLOPS FP8.

Directly benchmarked against Nvidia. Huawei claims Atlas 950 delivers 6.7x the compute and 15x the memory capacity of Nvidia's next-gen NVL144 supernode. Under tightening US export restrictions, this is the most aggressive performance label Huawei has ever slapped on a domestic alternative.

Q4 2026 production target. Huawei plans to begin shipments in Q4 2026, paired with in-house HBM. From its MWC 2026 debut to a physical demo at WAIC, the productization timeline is accelerating.

WangDou's Take

6.7x NVL144? That number is more useful on a slide deck than an engineering spec sheet. Huawei is quoting peak theoretical compute, but real-world training efficiency depends on interconnect bandwidth, compiler maturity, and software ecosystem — precisely the three areas where the Ascend stack is weakest. NVL144 has fifteen years of CUDA developer inertia behind it; Atlas 950 has CANN and MindSpore still chasing PyTorch. That said, under export controls that have all but cut off high-end GPU imports, Huawei doesn't need to win the spec war. It needs to be "usable and available." If Q4 shipments actually materialize, China's top foundation model labs won't be looking at Atlas 950 as the optimal choice — they'll be looking at it as the only choice.

Source: Huawei Central, Tom's Hardware